ONNX

ONNX gives trained models a shared file format, so a network built in one framework can run on another runtime or compiler. The spec has three pieces: an intermediate representation for computation graphs, a versioned catalog of standard operators grouped into opsets, and a list of tensor data types, all serialized with Protocol Buffers. It targets inference rather than training and covers classic machine learning as well as deep networks. The onnx package from pip is a toolkit for the format, not an execution engine: it builds and validates graphs, infers tensor shapes, converts models between opset versions, and runs operators through an optional pure Python reference backend for testing. Models usually arrive through exporters in PyTorch, TensorFlow and Keras, scikit-learn, or Hugging Face Optimum, then run on engines such as ONNX Runtime; ready-made models live in the onnx-community space on Hugging Face. An open steering committee and SIGs under the Linux Foundation govern it, the code is Apache 2.0, and releases land several times a year. Core knowledge for anyone moving models between stacks.